On Measuring Influence in Avoiding Undesired Future

ICLR 2026 Conference Submission18687 Authors

Keywords: avoiding undesired future, rehearsal learning, causal learning

TL;DR: We provide a principled framework for measuring influence in avoiding undesired future.

Abstract: When a predictive model anticipates an undesired future event, a question arises: what can we do to avoid it? The key to resolving this forward-looking challenge lies in determining the right variables that influence the result, moving beyond statistical correlations typically exploited for prediction. In this paper, we introduce a novel framework for evaluating the influence of alterable variables in successfully avoiding the undesired future. We quantify influence as the degree to which the probability of success can be increased by altering variables based on the principle of maximum expected utility. A crucial insight from our analysis is that the most influential variables may not necessarily be those with inherently strong causal effects on the future event. In fact, it can be highly beneficial to alter a weak causal ancestor, or even a variable that is not a causal ancestor at all. Furthermore, to overcome the practical challenges of exact computation, we provide a Monte-Carlo method for efficiently assessing influence using observational data. Experiments demonstrate the empirical performance of the proposed framework.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 18687